def _calculate_measure()

in domainbed_measures/measures/classical.py [0:0]

• 21 lines of code
• 7 McCabe index (conditional complexity)

    def _calculate_measure(self, sigma_max=2.0, sigma_min=0.0):
        """
        Compute the sharpness magnitude 1/alpha'^2 described in [1].

        Notes
        -----
        - This is slightly different than [1] because the target deviation is
        on cross-entropy instead of accuracy

        Args:
            sigma_max: float, optional
            sigma_min: float, optional
                Minimum standard deviation of perturbation.
        """
        trainer = clone_trainer(self._trainer_current)
        trainer.criterion = self._measure_criterion
        trainer.initialize()

        acc = self.accuracy_trainer(trainer, self._train_loader)
        logging.info(f"Accuracy of original model: {acc}")

        for bin_search in range(self._max_binary_search):
            sigma_min, sigma_max = self.get_sharp_mag_interval(
                trainer,
                acc,
                sigma_min,
                sigma_max,
            )

            if sigma_min > sigma_max or math.isclose(
                    sigma_min, sigma_max, rel_tol=1e-2):
                # if interval for binary search is very small stop
                break

        if bin_search == self._max_binary_search - 1:
            logging.info(
                f"Stopped early beacuase reached max_binary_search={self._max_binary_search}.\
                [sigma_min,sigma_max]=[{sigma_min},{sigma_max}]")

        return 1 / (sigma_max**2), {}